Constrained Bootstrap Methods

• Constrained bootstrap is a resampling and optimization approach that integrates explicit constraints (e.g., symmetry, positivity) to ensure statistically and physically valid solutions.
• It utilizes linear, affine, and semidefinite programming to enforce conditions such as crossing symmetry and boundary restrictions in both theoretical physics and statistical inference.
• Empirical studies demonstrate that constrained bootstrap methods yield improved finite-sample performance, tighter bounds, and corrected asymptotic properties in diverse applications.

A constrained bootstrap refers to a broad class of methodological frameworks in which bootstrap procedures—optimization-based or resampling-based—are subject to explicit constraints reflecting either physical, statistical, or problem-specific structure. These constraints may stem from algebraic relations, positivity, symmetry, boundary conditions, submanifold structure, or null hypotheses with composite or boundary character. The constrained bootstrap paradigm features in theoretical physics (the conformal, S-matrix, and quantum many-body bootstraps), classical and modern statistics (order-restricted inference, hypothesis testing on submanifolds), and applied domains (distributionally robust optimization, time series of compositional data). The following sections synthesize the conceptual foundations, algorithmic realizations, and illustrative applications drawn from recent arXiv literature.

1. Mathematical Foundations and General Framework

At its core, the constrained bootstrap is an optimization problem over a set of variables (operator averages, parameters, moments, or amplitudes) whose feasible set is delineated by a collection of linear, affine, or positive-semidefiniteness constraints. In the most abstract form, the problem is

$\sup/\inf_{x \in \mathcal{C}} O(x)$

subject to

$A(x) = 0,\quad G_j(x)\geq 0,\quad M(x) \succeq 0,$

where $x$ represents the variables (such as operator expectation values, parameter vectors, or OPE coefficients), $A$ encodes linear constraints (crossing, symmetries, normalization), $G_j$ are inequality constraints (e.g., positivity, boundary conditions), and $M(x)$ is a moment or Gram matrix subjected to semidefinite constraints. These problems are amenable to linear or semidefinite programming (SDP or LP) solvers. The constraints reflect deep structure—symmetry, unitarity, reflection positivity, or problem-specific null hypotheses—ensuring that only physically/statistically admissible solutions are explored (Zheng, 2023).

2. Constrained Bootstrap in Theoretical and Mathematical Physics

The modern conformal and S-matrix bootstrap programs are prototypical examples of constrained bootstrap applications in physics (Zheng, 2023, Gliozzi, 2013, Han, 2020). Constraints derive from symmetry (OPE associativity/crossing symmetry in CFT, analyticity/unitarity in S-matrix), positivity (OPE coefficient squares, spectral densities), and physical selection rules. For quantum many-body systems, the constrained bootstrap formalism—specifically, the quantum many-body bootstrap—employs expectation values over local operators subjected to normalization, adjoint-reality, positivity, and symmetry constraints. The approach yields certified lower bounds on ground state energies directly in the thermodynamic limit and provides tight bounds for ground state observables such as double occupancy and magnetization. The feasible region for e.g. the Hubbard model is determined by operator locality, and optimization is cast as a finite-dimensional SDP (Han, 2020).

The quantum mechanical bootstrap on bounded domains extends the framework: positivity, operator algebra, and recursion relations among moments (reflecting the underlying commutators and Hamiltonians) define a finite constraint system wherein increasing matrix size rapidly converges to the exact eigenvalues of the problem (Sword et al., 2024).

3. Constrained Bootstrap in Statistical Inference

In statistics, constrained bootstrap methods explicitly enforce structure arising from boundary conditions, submanifold hypotheses, or order restrictions. The parametric bootstrap percentile approach for constrained (e.g. order- or boundary-restricted) parameters exhibits nonstandard asymptotic coverage properties near the boundary; in one-sample settings, the coverage is always conservative, while in two-sample order-restricted problems, under- or overcoverage can occur depending on placement near the constraint boundary (Wang et al., 2017). The essential idea is to generate bootstrap samples or p-values from constrained maximum likelihood estimators that respect the parameter restriction, leading to vacancy or mass concentration at the boundary in the resampling distribution.

The constrained bootstrap for hypothesis testing on submanifolds (e.g., matrix rank or composite nulls) operates by resampling under the null-constrained maximum likelihood estimator or projection, then recalculating the test statistic. This procedure, which can be viewed as a least-squares constrained estimation (LSCE) algorithm coupled to a bootstrap, provides improved small-sample accuracy and maintains asymptotic correctness for the desired level and power (Portier et al., 2013).

In high-dimensional or stratified inference, the constrained parametric bootstrap for signed likelihood root or profile-score pivots corrects for bias and achieves higher order accuracy than the unconstrained bootstrap, particularly when the number of nuisance parameters is large or diverges rapidly relative to sample size (Bellio et al., 2020).

Further, equivalence testing for multinomial distributions using non-differentiable distance metrics (such as the sup- or $L^1$-norm) benefits from a constrained bootstrap in which the MLE is projected onto the (possibly boundary) null region. This projection step is critical for asymptotic validity in the presence of composite or boundary nulls (Bastian et al., 2023).

4. Algorithmic Realizations and Problem-Specific Constraints

The practical implementation of constrained bootstrap techniques requires systematic translation of physical, algebraic, or statistical structure into explicit constraints in the optimization or resampling procedures. In physics-driven bootstraps, these constraints emerge as:

• Crossing symmetry (CFT): Linear equations equating different OPE channel decompositions, expanded in basis function derivatives or truncated operator spectra (Gliozzi, 2013, Beem et al., 2015).
• Unitarity/positivity: Non-negativity for OPE coefficients, moment or semi-infinite matrices, partial-wave amplitudes (Zheng, 2023).
• Symmetry and conservation: Projectors or linear functionals encoding global, discrete, or gauge symmetries, constraining the allowed operator basis or parameter space (Zheng, 2023, Han, 2020).

In statistical contexts, the core constraint arises either from boundary- or order-restricted parameter spaces, submanifold structure (e.g., matrix manifolds or locally smooth parameterizations), or equivalence regions in composite null testing. The usual resampling or parametric algorithms are modified to sample from constrained estimators or project the empirical estimates into the feasible set before resampling. For example, in the multinomial equivalence scenario, computation of the critical value requires projecting the unconstrained MLE into a boundary point on the null set if the observed difference falls within the null; the subsequent bootstrap replicates then respect the constraint, enabling asymptotically valid inference even for non-differentiable distances (Bastian et al., 2023).

A comprehensive summary of major problem classes and their associated constraint types appears below:

Domain	Type of Constraint	Realization
Quantum many-body (physics)	Operator moments, locality	SDP, moment matrix PSD
Conformal/S-matrix bootstrap	Crossing, positivity	LP/SDP, spectral bounds
Statistical inference	Boundary/order restrictions	Constrained MLE, resampling
High-dimensional/preferential	Submanifold/projection	LSCE, manifold optim.

5. Empirical Performance and Theoretical Guarantees

Constrained bootstrap procedures frequently display superior empirical performance over their unconstrained counterparts in the presence of structure or finite-sample effects. Examples include:

• Improved finite-sample accuracy: In matrix rank tests and locally smooth submanifold hypotheses, constrained bootstrapping preserves nominal level and power under composite nulls, correcting for the bias and miscalibration of asymptotic approximations. In extensive simulations, correct rejection frequencies are achieved even at small $n$ or high-dimensional settings (Portier et al., 2013, Bellio et al., 2020).
• Rigorous bounds in numerical bootstrap: In quantum many-body systems and CFTs, the constrained SDP bootstrap yields lower or upper bounds on critical quantities (such as ground state energies, OPE coefficients) that systematically improve with the cutoff, converging in many cases to the known exact values or saturating known duality bounds (Han, 2020, Sword et al., 2024, Beem et al., 2015).
• Asymptotic properties in boundary-inference: Constrained percentile intervals are always conservative in one-sample boundary scenarios but require caution in two-sample or multi-parameter cases, where undercoverage can occur. The local limit asymptotics precisely quantify the regime and magnitude of conservatism or anti-conservatism (Wang et al., 2017). For composite nulls with non-smooth norms, the constrained bootstrap achieves correct type I error and consistency even when classical bootstrap fails (Bastian et al., 2023).

6. Extensions, Generalizations, and Future Outlook

The constrained bootstrap paradigm is unified in that the core workflow—selecting an observable or test statistic, formulating all relevant constraints, and recasting as an optimization or constrained sampling procedure—applies across a wide spectrum of disciplines. Recent works generalize this architecture to distributionally robust optimization (DRO), in which bootstrap resampling underpins the construction of finite-sample ambiguity sets, leading to convex programs with explicit safety guarantees (Summers et al., 2021). In the analysis of compositional or constrained time series (e.g. life-table death counts), transformations to an unconstrained scale (such as the centered log-ratio transform), followed by constrained bootstrapping and back-transformation, enable uncertainty quantification while respecting non-negativity and sum constraints (Shang, 16 Jul 2025).

Open directions include the extension of constrained bootstrap to more general probabilistic, functional, or deep learning settings, integration with robust decision architectures, and algorithmic advances for larger-scale SDPs and manifold projections. The unifying perspective remains the systematic encoding of structured information as constraints within bootstrap-based inference or optimization pipelines.

---

References:

• Quantum Many-body Bootstrap (Han, 2020)
• More constraining conformal bootstrap (Gliozzi, 2013)
• Bootstrap Testing of the Rank of a Matrix via Least Squared Constrained Estimation (Portier et al., 2013)
• Bootstrap prediction intervals for the age distribution of life-table death counts (Shang, 16 Jul 2025)
• Distributionally Robust Bootstrap Optimization (Summers et al., 2021)
• Testing equivalence of multinomial distributions -- a constrained bootstrap approach (Bastian et al., 2023)
• Asymptotic coverage probabilities of bootstrap percentile confidence intervals for constrained parameters (Wang et al., 2017)
• Parametric bootstrap inference for stratified models with high-dimensional nuisance specifications (Bellio et al., 2020)
• Quantum mechanical bootstrap on the interval: obtaining the exact spectrum (Sword et al., 2024)
• Bootstrap Method in Theoretical Physics (Zheng, 2023)
• Soft Bootstrap and Supersymmetry (Elvang et al., 2018)
• The $(2,0)$ superconformal bootstrap (Beem et al., 2015)
• Bootstrap percolation, probabilistic cellular automata and sharpness (Hartarsky, 2021)
• The Boostless Bootstrap and BCFW Momentum Shifts (Stefanyszyn et al., 2020)